Frequency-aligned supervision for few-shot neural rendering

Abstract

Neural rendering has shown significant potential in generating high-quality 3D scenes from sparse inputs. However, existing methods struggle to simultaneously capture both low-frequency global structures and high-frequency fine details, leading to suboptimal scene representations. To overcome this limitation, we propose a frequency-aligned supervision framework that explicitly separates the learning process into low-frequency and full-spectrum components. By introducing two sub-networks and aligning supervision signals at appropriate layers, our method enhances the formation of global structures while preserving fine details. Specifically, the low-frequency network (LFN) is supervised with low-pass targets (Gaussian-filtered images) to form global structures, while the full-spectrum network (FSN) is supervised with the original images to refine high-frequency details. The proposed approach is broadly applicable to MLP-based NeRF architectures without requiring major architectural modifications. Extensive experiments demonstrate that our method consistently improves PSNR, SSIM, and LPIPS across multiple NeRF variants and datasets, confirming its robustness in sparse input scenarios.